Embodied AI

Embodied Artificial Intelligence
Research and development of artificial agents whose perception, actions, and behavior involve interaction with physical or simulated environments.
Category	Artificial intelligence research field
Area	Artificial Intelligence, Robotics, Machine Learning, Cognitive Science, Control Theory
Applications	Mobile robotics, Robotic manipulation, Human–robot interaction
Related concepts	Physical AI, Autonomous Robots, Robot Foundation Models, Sensor Fusion, Embodied Cognition

[[File:Spot by Boston Dynamics.jpg |thumb |right |280px |alt=A four-legged mobile robot developed by Boston Dynamics. |Spot, a quadrupedal mobile robot developed by Boston Dynamics. Photograph: Jonte, licensed under CC BY-SA 4.0. Image source. ]] Embodied artificial intelligence (embodied AI) is an area of artificial intelligence concerned with the development and study of intelligent agents whose capabilities involve interactions between their bodies, control systems and environments. Rather than treating intelligence solely as the processing of abstract information, embodied AI investigates how perception, action, physical or simulated embodiment, and environmental feedback contribute to intelligent behavior.

Embodied agents may take the form of physical robots or virtual agents operating in simulated environments. Their interactions may involve navigation, object manipulation, instruction following, active perception and other tasks requiring actions that influence subsequent observations. Simulation-based embodied AI research includes the development of environments in which agents can be trained and evaluated under controlled conditions.

The concept is related to embodied cognition, a group of approaches in cognitive science that investigate relationships between cognition, bodily processes and environmental interaction. However, embodied cognition encompasses several distinct theoretical positions, while embodied AI primarily concerns the construction, modeling and evaluation of artificial agents.

Embodied AI is not restricted to humanoid robots or to a particular machine-learning architecture. Different research systems may employ conventional control methods, explicit world models, reinforcement learning, imitation learning or combinations of these approaches.

Historical and conceptual background

Situated artificial intelligence

[File:Rethink Robotics — Brooks and Baxter (8000143255).jpg |thumb |right |290px |alt=Robotics researcher Rodney Brooks standing beside a Baxter robot. |Roboticist Rodney Brooks with the Baxter robot developed by Rethink Robotics. Photograph: Steve Jurvetson, CC BY 2.0. Image source. ] The development of embodied AI is closely related to research on situated intelligence and behavior-based robotics.

In his 1991 paper "Intelligence Without Representation", roboticist Rodney Brooks argued for an approach in which intelligent behavior could be developed incrementally through direct interactions between a robot and its environment. His proposed architecture emphasized parallel behavioral processes connected to perception and action rather than requiring all behavior to depend on a centralized symbolic representation of the world.

Brooks's work contributed to the development of behavior-based approaches to robotics. It also provided an influential alternative to artificial-intelligence architectures that placed greater emphasis on centralized planning and symbolic representations.

These approaches did not establish that internal representations are unnecessary for every intelligent system. Instead, they demonstrated that direct environmental interaction and relatively simple behavioral components could produce useful robotic behavior under particular conditions.

Embodied intelligence

Research on embodied intelligence examines how an agent's physical characteristics contribute to its behavior and control.

Pfeifer and colleagues emphasized that the design of an intelligent agent involves relationships among its control system, body morphology, material properties and environment.

This perspective includes the concept of morphological computation, in which physical structures and material properties contribute to functions that might otherwise require more explicit computational control.

Embodied intelligence is therefore concerned not only with software and algorithms but also with the ways in which the characteristics of an agent's body influence its possible behaviors.

Fundamental principles

Perception–action interaction

A central feature of embodied AI is the relationship between perception and action.

An embodied agent receives observations through sensors or simulated observation systems, processes information relevant to a task, and selects actions that affect its environment. The resulting changes may influence subsequent observations and decisions.

For a physical robot, relevant sensory information may include camera images, depth measurements, joint positions and other measurements of the robot or its surroundings. In simulated environments, comparable observations can be generated by the simulator.

This continuing interaction is commonly described as a perception–action loop.

The loop differs from a purely passive perception task because the agent's actions can change what it subsequently observes.

Embodiment and morphology

The body of an intelligent agent influences how it can interact with its environment.

Physical characteristics such as dimensions, mechanical structure, degrees of freedom and material properties constrain the actions available to a robot. They can also facilitate particular behaviors.

Research on morphological computation investigates how mechanical design can contribute to adaptive behavior and reduce some demands placed on explicit control processes.

Consequently, embodiment is not simply the attachment of an artificial-intelligence model to a mechanical device. The relationship between the agent's body, its control system and its surroundings is part of the research problem itself.

Environmental interaction

Embodied AI studies agents that perform tasks through interactions with their surroundings.

A 2025 survey of embodied AI research organized robot behavior into three broad components: embodied perception, embodied decision-making and embodied execution. The authors used this classification to examine methods and challenges associated with robot mobility and manipulation.

These components are analytically distinguishable but interconnected. Actions may depend on sensory information, while decisions about movement and manipulation affect the information available to the agent.

Research tasks and applications

Navigation

Embodied navigation concerns the movement of an agent through an environment toward a destination or according to specified instructions.

Research may involve recognizing relevant surroundings, selecting routes, avoiding obstacles and responding to changes in the environment.

Navigation systems can be studied using physical mobile robots or simulated agents operating in three-dimensional environments.

An important research area is vision-and-language navigation, in which an agent attempts to follow natural-language navigation instructions using visual observations.

Anderson and colleagues introduced the Room-to-Room benchmark in 2018 to study navigation according to natural-language instructions in environments represented using imagery from real buildings.

Object manipulation

Object manipulation involves actions intended to change the position, orientation or physical state of objects.

Research tasks include grasping, moving and placing objects, as well as performing sequences of actions involving tools or other physical objects.

Manipulation requires coordination between the agent's perception of its surroundings and the execution of actions through its available mechanisms.

Machine-learning approaches to manipulation include learning from demonstrated examples and learning through interaction with an environment.

Instruction following

Instruction-following research examines how an agent interprets task descriptions and translates them into actions.

Instructions may concern navigation, manipulation or sequences of operations.

In vision-and-language navigation, instructions are interpreted in relation to visual observations and the agent's possible movements.

Research involving multimodal models has also investigated connections among language, visual observations, robot-state information and robotic task planning.

Interactive perception

Interactive perception concerns the use of actions to obtain information relevant to perception.

Unlike passive image recognition, an embodied agent may be able to change its position or viewpoint during a task.

Embodied navigation environments provide examples of systems in which the observations available to an agent depend on its movements through an environment.

Learning and control methods

Learning from demonstration

Learning from demonstration, sometimes called imitation learning in relevant settings, involves developing a policy or behavior from examples of task performance.

Demonstrations may be provided through teleoperation, observation or other forms of human or machine-generated examples.

Argall and colleagues surveyed methods for collecting demonstrations, deriving control policies and evaluating learned behavior.

Reinforcement learning

Reinforcement learning provides a framework in which an agent improves a policy using feedback associated with its actions.

In robotics, reinforcement-learning methods have been studied for control and behavior generation.

Kober, Bagnell and Peters reviewed both the potential and practical challenges of reinforcement learning for robotics, including the complexity of robotic systems and the importance of suitable representations and prior knowledge.

Planning and conventional control

Embodied AI research is not restricted to machine-learning methods.

Research platforms such as Habitat support comparisons involving learned policies and conventional navigation approaches, including simultaneous localization and mapping.

Such comparisons allow researchers to investigate different approaches to navigation and other embodied tasks under specified experimental conditions.

Simulation and transfer to physical environments

Simulation environments

Simulation is an important research method in embodied AI.

A simulator can provide virtual environments, artificial sensor observations and defined interactions between agents and their surroundings.

Researchers can use these environments to train agents, conduct repeated experiments and compare methods under controlled conditions.

The Habitat platform, introduced by Savva and colleagues in 2019, was developed for research on embodied agents in three-dimensional environments.

It includes Habitat-Sim, a simulation system, and Habitat-API, a framework supporting task definition, agent training and evaluation.

The platform was designed to support research tasks such as navigation, instruction following and embodied question answering.

The simulation-to-reality gap

A major difficulty in transferring systems from simulation to physical hardware is the difference between simulated conditions and real-world environments.

These differences may affect the performance of perception and control systems.

Research on simulation-to-real-world transfer, commonly called sim-to-real transfer, investigates methods for improving the ability of systems developed in simulation to operate in physical environments.

Tobin and colleagues investigated domain randomization, in which simulated visual conditions are varied during training to improve transfer to real-world observations.

Their 2017 work demonstrated the use of simulated training data for object localization relevant to robotic manipulation.

Success in a simulated environment does not, by itself, establish equivalent performance on physical hardware. Evaluation of practical robotic capabilities therefore requires attention to the conditions under which a system is tested.

Embodied AI and foundation models

Research has investigated the incorporation of large pretrained models into systems that process embodied observations or produce actions.

These approaches include multimodal language models, robotic policies and models that combine vision, language and action representations.

Embodied multimodal language models

Driess and colleagues introduced PaLM-E in 2023.

PaLM-E combines language processing with inputs representing visual information and continuous state measurements.

The researchers investigated its use for tasks including robotic manipulation planning, visual question answering and image captioning.

The work illustrates an approach to connecting language-based models with information relevant to embodied tasks.

Vision-language-action models

Vision-language-action models (VLAs) are models designed to connect visual observations, language-related information and action generation.

RT-2, reported by Zitkovich and colleagues in 2023, adapted vision-language models for robotic control by representing robotic actions as tokens within the model's output.

The researchers combined robotic trajectory data with vision-language training data and evaluated the resulting system on robotic tasks.

RT-2 is an example of an architecture connecting pretrained vision-language capabilities with robotic action prediction. Its reported results concern the experimental systems and tasks described in the research publication.

Learning across robot platforms

Researchers have also examined whether models can learn from data collected using different robotic systems.

The Open X-Embodiment Collaboration investigated the use of datasets gathered across multiple robot embodiments and developed RT-X models for robotic learning.

The work was reported at the 2024 IEEE International Conference on Robotics and Automation.

Such research addresses the challenge of developing models that can make use of experience from more than one robotic platform. The performance of a learned model remains dependent on its training conditions, target embodiment and evaluation tasks.

Evaluation and research limitations

Performance evaluation

Embodied AI systems are evaluated according to the tasks and environments for which they are designed.

Navigation research commonly examines whether an agent reaches a specified destination and how effectively it performs the task.

Manipulation research may examine task completion, successful object handling and performance across different experimental conditions.

Researchers may also investigate how well an agent performs when its environment or task conditions differ from those encountered during training.

Habitat was developed partly to facilitate repeatable benchmarking and comparisons between embodied AI methods.

Generalization

Generalization concerns the performance of a system under conditions that differ from those represented in its training data.

For embodied agents, relevant differences may involve environments, objects, tasks or robotic platforms.

Research on vision-language-action models and cross-embodiment learning has investigated particular aspects of generalization using defined experimental procedures.

Performance on one benchmark does not establish equivalent capabilities across all possible environments or embodiments.

Physical safety

Embodied systems operating through physical robots may introduce mechanical or operational hazards.

Requirements for safety depend on the type of robot, its application and the environment in which it operates.

For industrial robotics, ISO 10218-1:2025 specifies safety requirements for industrial robots, while ISO 10218-2:2025 addresses industrial robot applications and robot cells.

These standards have defined scopes and do not establish universal safety requirements for every category of embodied AI system.

The development of an embodied AI model and the demonstration of performance on selected research tasks should therefore be distinguished from establishing the safety of an integrated physical system.

Relationship to other fields

Robotics

Robotics concerns the design, construction, control and operation of robots.

Embodied AI overlaps with robotics when artificial-intelligence methods are used to investigate perception, decision-making and action in robotic systems.

Not every aspect of robotics is necessarily an embodied AI research problem, and embodied AI research can also be conducted with simulated agents.

Embodied cognition

Embodied cognition is an interdisciplinary field concerned with relationships between cognitive processes, bodily activity and environmental interaction.

Wilson's review identified several distinct claims associated with embodied cognition and emphasized that they differ in their theoretical implications and evidential support.

Embodied AI draws on related ideas while focusing on artificial agents, their designs and experimentally observable behavior.

Physical AI

Embodied AI is closely related to the application of artificial intelligence in physical systems.

In robotics, both concern the relationship between computational processes, perception and action.

The terms are not necessarily interchangeable across all research contexts, particularly because embodied AI research can employ virtual agents in simulated environments.

See also

• Physical AI
• Robotics
• Autonomous Systems
• Autonomous Robots
• Computer Vision
• Sensor Fusion
• Reinforcement Learning
• Embodied Cognition
• Robot Learning
• Sim-to-Real Transfer
• Robot Foundation Models
• Robot Manipulation
• Autonomous Navigation